TXT1a-1 (v2): the text-latin registry against Chrome, engine-mode real faces and R3 vectors

[thejackshelton]

TXT1a-1 (v2): the text-latin registry against Chrome 145, with engine-mode lowering of bundled static real faces. It is the third level of the text stack. #101 and #102 (txt1a-1b-v2) are on master, so this PR's base is now master.

What changed

• Compiler (packages/dragon).
  • lower/ios-layout.ts: lowerTree takes a lowering mode. Native lays out Ahem only. Engine mode also lays out the bundled static face that each element's and text node's font resolves to (face id sha256:<hex>).
  • project.ts:
    • It records the native font refusals the engine lane defers (fontDeferrals).
    • When every native blocker is a deferred font refusal, it lowers the native band's cases in engine mode (engineFacesOf, engineFaceFor).
    • Engine mode refuses fallback to several faces, variable faces and synthetic styles.
    • DRAGON_SYNTHETIC_FONT_STYLE (codes/txt1a.ts) refuses native text that Chrome would draw in synthetic bold or oblique.
  • internal.ts: engineLayoutProjection.
• Parity.
  • text-latin-run.ts and fixture-groups/text-latin.ts: the text-latin registry (6 fixtures, 7 cases, plus 3 probes). It is kept out of FIXTURES, and its GROUPS entry is empty.
  • Every case must pass all of these:
    • the engine lane, exact at 1/64 px with the Node HarfBuzz host;
    • chrome-dual;
    • the face check (CSS.getPlatformFontsForNode) in both documents;
    • the break check against Chrome, at DPR 1, 2, 3 and 2.625.
  • runCase and runDprCase take the projection as an argument.
  • referenceShapedMeasurer registers every vendored face.
  • report.ts and gen-profile-rows.ts run the registry. Its cases prove web font-family rows only.
• Vectors (R3). packages/layout/vectors/text-latin/dpr-<d>/<case>.json holds 28 vectors: the engine input and output with the shape transcript (faces and every HarfBuzz call).
• Merge with TXT1a-1b (v2): real-font Latin through HarfBuzz in the engine, restacked onto INL1a C2 #102's head (d542a45) and with master. Conflicts in ios-layout.ts and project.ts were resolved keeping both sides:
  • the Lowerer carries both the mode and master's propagated;
  • checkCases takes the @Property registrations and keeps the font deferrals beside checkNativeScroll;
  • native lowering is master's per-band loop in native mode, then the engine-mode lowering.
• Fixes after the merge and the audit.
  • line-breaks.ts engineTextLines re-resolved its input with zoomInput (Ahem) after laying it out with the caller's measurer. It now uses resolvedInput with the same measurer, so a real face's ex, ch, cap or lh no longer throws text-glyph. resolvedInput is exported from @dragon/layout.
  • cli/text-latin-capture.ts:
    • An unknown argument exits 2 before anything is cleaned. A mistyped --vectors used to run the live capture, which first deletes the committed captures.
    • A case that writes nothing exits 1 instead of only logging.
  • Two new regen steps in scripts/regen-steps/txt1a.ts, each with its .gitattributes line. They replace the old MANUAL entry, which a cloud lane cannot run:
    • text-latin-capture runs after capture, on the macos-26 capture runner. It writes the live Chrome captures, breaks and emitted CSS of the registry, as capture does for expected-fonts and expected-env. The committed captures predated master's new computed keys (border--radius, content, list-style-), so every text-latin case failed chrome-dual in profile:rows and text-latin-faces lost its web output.
    • text-latin-vectors runs after dpr-vectors. It rebuilds the vectors from those captures. The vectors predated master's grid keys.
  • The generated Swift and Kotlin replays of the vectors move from text-latin.test.ts to text-latin-native.test.ts, unchanged. The old file needed both Chrome and the native toolchains, so it could pass on neither CI shard.

Outside the spec

• scripts/regen-steps/txt1a.ts, .gitattributes: the text-latin-capture and text-latin-vectors regen steps. The captures and vectors predated master's computed keys and grid keys, and had to be rebuilt by regen.
• packages/parity/test/text-latin-native.test.ts (new): the native-shard half of text-latin.test.
• packages/layout/src/index.ts: exports resolvedInput.
• packages/parity/src/line-breaks.ts: engineTextLines resolves with the caller's measurer.
• packages/layout/test/inline-box-placelines.test.ts: see below.

Tests and pins changed, with reasons

• packages/parity/test/text-latin.test.ts:
  • BASE pin moves from 569 to 666 layout cases, and the id hash from 3c0c5d0b… to caefcdb7…. These are master's FIXTURES at b97cf50, after TXT1a-1b (v2): real-font Latin through HarfBuzz in the engine, restacked onto INL1a C2 #102, ANIM-b1 native runtime 1/2: animation tables as typed literals, the translated animator on the state machine, display drivers #203, GEN-c: list item ordinals and marker text (counter styles), compile-time #232 and CASC 3: @layer decided at build time as Chrome 145 orders cascade layers #233; the ids were checked identical on an origin/master worktree. The intent is kept: TXT1a-1 changes no FIXTURES case.
  • The Swift and Kotlin replays move to text-latin-native.test.ts, unchanged, which also checks that it reads all 28 vectors.
  • Adds a test that the capture CLI refuses an unknown argument with exit 2.
• packages/parity/test/text-latin-engine.test.ts: referenceShapedMeasurer() now holds every vendored face, Lato included. So "a shaped measurer without the face" is now an explicit Ahem-only shaped measurer (ahemShapedMeasurer). A new test pins that this measurer lacks Lato and that the reference one holds it. Each refusal is still judged on a measurer without the face.
  • Adds a test that engineTextLines resolves a Lato ex width. It fails with the old zoomInput code.
• packages/layout/test/inline-box-placelines.test.ts: the Ahem measurer refused the text-latin vectors on both sides of placeLines == placeIfcLines, so the check passed without placing a line. Those vectors are now measured by the HarfBuzz reference measurer, every placement must succeed, and the 28 shaped vectors are counted. Forcing Ahem makes the test fail.
• packages/parity/test/parity.test.ts: the fixture-file existence check now covers the text-latin fixtures and probes, not only the font and env ones.
• packages/layout/test/dpr-vectors.test.ts: the vectors folder also holds text-latin.

What passed (head fafd726)

• Regen on CI: pnpm ci:regen txt1a-1-v2, https://github.com/compiled-run/dragoncss/actions/runs/37931754092. Its regen commit is 4ff2514, on top of the master merge 6bc0e63. The only later commit (fafd726) changes test files only, which are not generator inputs.

• pnpm typecheck (local): pass.

• Local platform-free runs:

  • packages/layout/test: 47/47 files, 1444 tests pass. This includes inline-box-placelines on the regenerated text-latin vectors.
  • packages/dragon/test: 75 of 76 files pass. The exception is native-backends, which needs the native toolchains and passed on CI below.

• pnpm ci:test-files at fafd726, https://github.com/compiled-run/dragoncss/actions/runs/37944523853: PASSED, all 39 files with 0 failures:

  • Native shard: text-latin-native (the Swift and Kotlin vector replay), regen, native-swift, native-kotlin, native-dpr-swift, native-dpr-kotlin and native-backends.
  • Chrome shard, targeted files: text-latin, text-latin-engine, line-breaks, parity (556), translate (both), chrome-ports, parity registry-claims and android-profile.
  • Chrome shard, fixture-wide set: p6a-dash-oracle, native-host, device-lanes, dist, env-engine, fixture-reader, fonts-fixtures, inline, interaction-forced, media-runtime, media-sweep, phrasing-blockified, platform, replaced-identity, tree-faults, values, glyph-calibration, hit-report, hit-capture-identity, pixel-reference, pnt2-quads, pnt2-samples and replaced-paint.

• pnpm -s pr:review 103 --once: exit 0. CI checks pass with 0 unanswered findings. The head is UNREVIEWED because Macroscope hit its monthly spending limit.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WSkSiXMPXTAsNpS9BD9LDR

---

Generated by Claude Code